```html

Artificial Intelligence Takes Off: CognitiveDrone Enables Cognitive Problem Solving for Drones

The rapid development in the field of Artificial Intelligence (AI) is constantly opening up new application possibilities. A particularly exciting field is the control of drones using AI, which goes beyond simple flight maneuvers and enables more complex, cognitive tasks. A research team has now presented "CognitiveDrone," a new Vision-Language-Action (VLA) model that enables drones to act and make decisions in real-time based on visual and linguistic information.

From Perception to Action: How CognitiveDrone Works

CognitiveDrone is based on an innovative approach that combines visual information from the drone's perspective with textual instructions. The model analyzes the surrounding images and interprets the given commands to derive corresponding actions. These actions are output as 4D control commands that control the drone in real-time. The model was trained using a comprehensive dataset with over 8,000 simulated flight trajectories. The tasks within the dataset can be divided into three main categories: human detection, symbol understanding, and logical reasoning.

CognitiveDrone-R1: Reasoning Power for Complex Scenarios

For particularly demanding tasks, an extended version of the model, CognitiveDrone-R1, was developed. This integrates an additional Vision-Language Model (VLM) as a reasoning module. This module simplifies complex instructions before they are passed on to the drone's control system. This intermediate step allows the drone to act more effectively even in complex scenarios and to execute the given tasks more precisely.

CognitiveDroneBench: A New Benchmark for Cognitive Drone Tasks

To evaluate the performance of CognitiveDrone, an open-source benchmark called CognitiveDroneBench was developed. This benchmark allows for direct comparison with other models and serves as a basis for further research in this area. Compared to a model trained on racing scenarios (RaceVLA) with a success rate of 31.3%, the basic model of CognitiveDrone achieved a success rate of 59.6%. The extended version, CognitiveDrone-R1, even achieved a success rate of 77.2%. These results highlight the potential of cognitive abilities in drone control and show that significant performance improvements, particularly in complex tasks, can be achieved through the integration of reasoning modules.

Outlook: Cognitive Drones in Action

The development of CognitiveDrone and CognitiveDroneBench represents an important step towards autonomous drones that can handle complex tasks in various application areas. From search and rescue operations to infrastructure inspections and precision agriculture – the possibilities are diverse. The combination of visual perception, language comprehension, and the ability to act opens up new perspectives for the use of drones in the future.

Bibliography: - https://arxiv.org/abs/2503.01378 - https://arxiv.org/html/2503.01378v1 - http://paperreading.club/page?id=288916 - https://github.com/jonyzhang2023/awesome-embodied-vla-va-vln - https://vlabench.github.io/ - https://cogact.github.io/ - https://openreview.net/forum?id=gkDRrvqeWF - https://paperreading.club/category?cate=Language_Model&page=3 - https://www.researchgate.net/publication/273297041_A_Cognitive_Task_Analysis_to_Elicit_Preliminary_Requirements_for_an_Automated_UAV_Verification_Planning_System ```